SJK Example

Transcript

1. ALIGN TECHNOLOGY, INC 1 © 2018 ALIGN TECHNOLOGY, INC 1 SJK example Dmitry Pavlov
2. ALIGN TECHNOLOGY, INC 2 Problem • Java server side application running in Docker • Growth of response metrics for one of the popular requests it handles • Higher CPU usage compared to daily stats • Existing metrics does not have “per class” precision, hard to identify root cause 99% 95%
3. ALIGN TECHNOLOGY, INC 3 Why SJK • Does not require any special jvm options • Works in Docker (compared to async-profiler by A. Pangin) • Way better than with further manual analysis and transfer back to dev machine for i in {1..15000}; do jstack 1 > $i.stack done
4. ALIGN TECHNOLOGY, INC 4 Collection and analysis • Ttop just to take a look. It shown cpu usage in tomcat threads • Time to sample stacks: • Use --thread-info and --histo for rough overview what do we collected and how to filter out irrelevant info • Build flame graph and meditate java -jar sjk.jar stcap -p 1 -o dump.std -t 60s
5. ALIGN TECHNOLOGY, INC 5 Findings • Ineffective RSA signature calculation on jdk8: https://bugs.openjdk.java.net/browse/JDK-8149554 - Verified with JMH and indeed it’s our case (app does a number of signatures per request) - Fixed with JDK feature flags that enable intrinsics in BigDecimal: -XX:+UseMontgomerySquareIntrinsic -XX:+UseMontgomeryMultiplyIntrinsic • Infrastructure layout issues